Page 424 - Sensing, Intelligence, Motion : How Robots and Humans Move in an Unstructured World

SALIENT CHARACTERISTICS OF A SENSITIVE SKIN 399

it. Ultrasound sensors can do this measurement easily, but their resolution is
not good.
One possible strategy is to adhere to a binary “yes–no” measurement. In a
sensor with limited sensitivity range, say 20 cm, the “yes” signal will tell the
robot that at the time of detection the object was at a distance of 20 cm from
the robot body. The technique can be improved by replacing a single sensor by
a small cluster of sensors, with each sensor in the cluster adjusted to a different
turn-on sensitivity range. The cluster will then provide a crude measurement of
distance to the object.

Sensors’ Physical Principle of Action. Vision sensing being as powerful as
we know it, it is tempting to think of vision as the best candidate for the robot
whole-body sensing. The following discussion shows that this is not so: Vision
is very useful, but not universally so. Here are two practical rules of thumb:

1. When the size of the workspace in which the robot operates is significantly
larger than the robot’s own dimensions—as, for example, in the case of
mobile robot vehicles—vision (or a laser ranger sensor) is very useful for
motion planning.
2. When the size of the robot workspace is comparable to the robot dimen-
sions—as in the case of robot arm manipulators—proximal sensing other
than vision will play the primary role. Vision may be useful as well—for
example, for the task execution by the arm end effector.

Let us start with mobile robot vehicles. When planning its path, a mobile
robot’s motion control unit will benefit from seeing relatively far in the direction
of intended motion. If the robot is, say, about a meter in diameter and standing
about a meter tall, with sensors on its top, seeing the scene at 10–20 meters
would be both practical and useful for motion planning. Vision is perfect for
that: Similar to the use of human vision, a single camera or, better, a two-camera
stereo pair will provide enough information for motion planning. On the other
hand, remember, the full coverage requirement prescribes an ability to foresee
potential collisions at every point of the robot body, at all times. If the mobile
robot moves in a scene with many small obstacles, possibly occluding each other
and possibly not visible from afar, so that they can appear underneath and at the
sides, even a few additional cameras would not suffice to notice those details.
The need for sensing in the vicinity of the robot becomes even stronger for
arm manipulators. The reason is simple: Since the arm’s base is fixed, it can reach
only a limited volume defined by its own dimensions. Thinking of vision as a
candidate, where would we attach vision cameras to guarantee the full coverage?
Should they be attached to the robot, or put on the walls of the robot work cell,
or both?
A simple drawing would show that in any of these options even a large number
of cameras—which is impractical anyway—would not guarantee the full sensing
coverage. Occlusion of one robot link by another link, or by cables that carry